What we're shipping and publishing.
Dated announcements, press coverage and papers. For the long-form guides and analysis, see Articles.
Compile-Time Inference launches
Verificate launches Compile-Time Inference — the alternative to RAG. Your facts are compiled, ranked and verified into a cited decision substrate before any question is asked; the model only phrases the answer. Three deployments at launch: Legal (Kevin), Video (Hybrid-VLM) and Genomics.
ReadBrown University — Sept 24–25
Verificate presents Compile-Time Inference at Brown University, Providence: "Intervention Bias as Disparate Impact" — how a decision policy compiled from an institution's own outcomes removes the over-intervention a zero-shot frontier agent introduces.
ReadVerificate presents Compile-Time Inference at AI Infra Summit, Santa Clara
Sept 15–17, Santa Clara Convention Center. The launch venue for Compile-Time Inference: live Kevin legal, Hybrid-VLM video and genomics demonstrations, with every answer resolving to its source.
ReadVividas presents Compile-Time Inference at IBC, Amsterdam
Partner Vividas presents the video deployment of Compile-Time Inference at IBC (International Broadcasting Convention), Amsterdam: a live broadcast feed compiled into a frame-stamped, provenance-linked record that a newsroom can query and verify.
ReadVerificate EAV-DT paper on arXiv
The deterministic decision-transformer substrate paper is released.
ReadVerificate v1.7 launched at UNSW
HELIX v1.7 (Build #58) goes live at UNSW — audited on the official vLLM GuideLLM benchmark: 940 ms p50 TTFT, 100% valid JSON, and 8.95–12.89× faster than stock Ollama.
UNSW becomes a foundation customer
After a successful pilot on its Intel-based Katana HPC, UNSW Sydney adopts Verificate HELIX — our flagship sovereign-inference deployment.
Verificate v1.4 launched for the UNSW trial
HELIX v1.4 is deployed for the UNSW evaluation on Intel HPC.
UNSW trial begins
The University of New South Wales starts a pilot of HELIX CPU inference on its Katana HPC.
Hallucination-prevention paper on arXiv (CORTEX)
Our inference-level hallucination-prevention method — a per-token truthfulness manifold — is released.
ReadIBM Business Partner established
Verificate becomes an IBM Business Partner, with access to IBM Fusion HPC infrastructure and Granite model integration.
The long-form version.
Why RAG still hallucinates — and what to change first
Your system finds the right documents and still gives wrong answers. How to tell retrieval failures from generation failures, why a citation can mislead, and the fixes worth trying before you change architecture.
ReadRAG vs Compile-Time Inference: what changes, what it costs, and when to use each
An honest side-by-side: where the work happens, who decides the answer, missing evidence, changing data and total cost — including where RAG remains the right choice.
ReadGraphRAG vs Compile-Time Inference: a graph connects information. What decides the answer?
What a knowledge graph improves, what it leaves unchanged, and the questions to ask any vendor — including us.
ReadPeer-facing research and audited technical reports. Measured results carry a M mark and are substantiated on our Legal & claims page.